Peeking into blackbox models

1. About myself and Go-jek
2. Interpreting Tabular (Gradient Boosting) Binary Classification
3. Interpreting Text/Document (Naive Bayes) Multi-Classification

Shameless self-promo

Zane Lim

• Senior data-scientist at Go-jek
• Lead and mentor a sub-team of data-scientists
• Mentor in Udacity's AI nanodegree
• Avid participant in Kaggle / data science hackathons

zyuanlim@gmail.com / zane.lim@go-jek.com

• https://www.linkedin.com/in/limzane/
• https://www.kaggle.com/zanelim

Go-Jek

• On-demand mobile platform that provides a variety of complete services
  • transportation (300,000 ojek and car driver partners)
  • logistics / food delivery (100,000 food vendors)
  • payment
  • other on-demand services (masseurs, cleaners etc.)

• As of July 2017, downloaded more than 44 million times
• Operates in 50 cities in Indonesia- Jakarta, Bandung, Surabaya, Bali, Makassar, Medan, Palembang, Semarang, Yogyakarta, Balikpapan, Malang, Solo, Manado, Samarinda, Batam, Sidoarjo, Gresik, Pekanbaru, Jambi, Sukabumi, Bandar Lampung, Padang, Pontianak, Banjarmasin, Mataram etc.

Drank

• Webservice that serves real-time driver rankings
• Based on predictions by machine learning model eg. probability of completion
• Rank drivers who are available in the vicinity of customer whenever a booking is requested
• Assign the job to the top ranked driver

Interpretation, Transparency and Fairness Evaluation

1. Driver's feedback- monitor drivers' complaints about not getting jobs
2. Promote transparency- explain and provide suggestions to drivers
3. Fairness / inequality- Gini coefficient of drivers income and jobs received
4. Measure average ranking / predicted probability of each driver
5. Model interpretability- Visualizing model, Partial Dependence Plot (PDP), Local Interpretable Model-Agnostic Explanations (LIME)

Interpreting Gradient Boosting Model

1. Direct Visualization
2. Features Importance
3. Partial Dependence Plot (PDP)
4. Local Interpretable Model-Agnostic Explanations (LIME)

import pandas as pd
import numpy as np
import lime
from lime import lime_tabular
import xgboost as xgb
import matplotlib.pyplot as plt
%matplotlib inline
import sklearn
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report
from sklearn.datasets import load_wine
from plotly.offline import download_plotlyjs, init_notebook_mode, plot, iplot
import plotly.graph_objs as go
init_notebook_mode(connected=True)

print 'pandas version: {}'.format(pd.__version__)
print 'numpy version: {}'.format(np.__version__)
print 'xgboost version: {}'.format(xgb.__version__)
print 'sklearn version: {}'.format(sklearn.__version__)

pandas version: 0.20.3
numpy version: 1.13.1
xgboost version: 0.6
sklearn version: 0.19.0

data = load_wine()

data['data'] = data.data[(data.target != 0)]
data['target'] = data.target[(data.target != 0)]-1
data['target_names'] = ['class_1', 'class_2']

print data.DESCR

Wine Data Database
====================

Notes
-----
Data Set Characteristics:
    :Number of Instances: 178 (50 in each of three classes)
    :Number of Attributes: 13 numeric, predictive attributes and the class
    :Attribute Information:
 		- 1) Alcohol
 		- 2) Malic acid
 		- 3) Ash
		- 4) Alcalinity of ash  
 		- 5) Magnesium
		- 6) Total phenols
 		- 7) Flavanoids
 		- 8) Nonflavanoid phenols
 		- 9) Proanthocyanins
		- 10)Color intensity
 		- 11)Hue
 		- 12)OD280/OD315 of diluted wines
 		- 13)Proline
        	- class:
                - class_0
                - class_1
                - class_2
		
    :Summary Statistics:
    
    ============================= ==== ===== ======= =====
                                   Min   Max   Mean     SD
    ============================= ==== ===== ======= =====
    Alcohol:                      11.0  14.8    13.0   0.8
    Malic Acid:                   0.74  5.80    2.34  1.12
    Ash:                          1.36  3.23    2.36  0.27
    Alcalinity of Ash:            10.6  30.0    19.5   3.3
    Magnesium:                    70.0 162.0    99.7  14.3
    Total Phenols:                0.98  3.88    2.29  0.63
    Flavanoids:                   0.34  5.08    2.03  1.00
    Nonflavanoid Phenols:         0.13  0.66    0.36  0.12
    Proanthocyanins:              0.41  3.58    1.59  0.57
    Colour Intensity:              1.3  13.0     5.1   2.3
    Hue:                          0.48  1.71    0.96  0.23
    OD280/OD315 of diluted wines: 1.27  4.00    2.61  0.71
    Proline:                       278  1680     746   315
    ============================= ==== ===== ======= =====

    :Missing Attribute Values: None
    :Class Distribution: class_0 (59), class_1 (71), class_2 (48)
    :Creator: R.A. Fisher
    :Donor: Michael Marshall (MARSHALL%PLU@io.arc.nasa.gov)
    :Date: July, 1988

This is a copy of UCI ML Wine recognition datasets.
https://archive.ics.uci.edu/ml/machine-learning-databases/wine/wine.data

The data is the results of a chemical analysis of wines grown in the same
region in Italy by three different cultivators. There are thirteen different
measurements taken for different constituents found in the three types of
wine.

Original Owners: 

Forina, M. et al, PARVUS - 
An Extendible Package for Data Exploration, Classification and Correlation. 
Institute of Pharmaceutical and Food Analysis and Technologies,
Via Brigata Salerno, 16147 Genoa, Italy.

Citation:

Lichman, M. (2013). UCI Machine Learning Repository
[http://archive.ics.uci.edu/ml]. Irvine, CA: University of California,
School of Information and Computer Science. 

References
----------
(1) 
S. Aeberhard, D. Coomans and O. de Vel, 
Comparison of Classifiers in High Dimensional Settings, 
Tech. Rep. no. 92-02, (1992), Dept. of Computer Science and Dept. of 
Mathematics and Statistics, James Cook University of North Queensland. 
(Also submitted to Technometrics). 

The data was used with many others for comparing various 
classifiers. The classes are separable, though only RDA 
has achieved 100% correct classification. 
(RDA : 100%, QDA 99.4%, LDA 98.9%, 1NN 96.1% (z-transformed data)) 
(All results using the leave-one-out technique) 

(2) 
S. Aeberhard, D. Coomans and O. de Vel, 
"THE CLASSIFICATION PERFORMANCE OF RDA" 
Tech. Rep. no. 92-01, (1992), Dept. of Computer Science and Dept. of 
Mathematics and Statistics, James Cook University of North Queensland. 
(Also submitted to Journal of Chemometrics). 

data.target_names

['class_1', 'class_2']

data.feature_names

['alcohol',
 'malic_acid',
 'ash',
 'alcalinity_of_ash',
 'magnesium',
 'total_phenols',
 'flavanoids',
 'nonflavanoid_phenols',
 'proanthocyanins',
 'color_intensity',
 'hue',
 'od280/od315_of_diluted_wines',
 'proline']

data.data.shape

(119, 13)

pd.Series(data.target).value_counts()

0    71
1    48
dtype: int64

X_train, X_test, y_train, y_test = train_test_split(data.data, data.target, test_size=0.3, random_state=0)

print X_train.shape, X_test.shape

(83, 13) (36, 13)

dtrain = xgb.DMatrix(X_train, y_train, feature_names=data.feature_names)

params = {'learning_rate': 0.1, 'max_depth': 4, 'subsample': 1.0, 'colsample_bytree': 0.8, 
          'objective': 'binary:logistic', 'eval_metric': 'logloss'}

cv = xgb.cv(params, dtrain, 200, nfold=5, stratified=True, early_stopping_rounds=20, seed=0)

cv.iloc[-1]

test-logloss-mean     0.072455
test-logloss-std      0.020414
train-logloss-mean    0.031961
train-logloss-std     0.000370
Name: 61, dtype: float64

params['n_estimators'] = 62
params['n_jobs'] = 5
params['random_state'] = 0
xgb_clf = xgb.XGBClassifier(**params)
xgb_clf.fit(X_train, y_train)

XGBClassifier(base_score=0.5, booster='gbtree', colsample_bylevel=1,
       colsample_bytree=0.8, eval_metric='logloss', gamma=0,
       learning_rate=0.1, max_delta_step=0, max_depth=4,
       min_child_weight=1, missing=None, n_estimators=62, n_jobs=5,
       nthread=None, objective='binary:logistic', random_state=0,
       reg_alpha=0, reg_lambda=1, scale_pos_weight=1, seed=None,
       silent=True, subsample=1.0)

y_probs = xgb_clf.predict_proba(X_test)
y_preds = xgb_clf.predict(X_test)

print classification_report(y_test, y_preds, target_names=data.target_names)

             precision    recall  f1-score   support

    class_1       1.00      0.91      0.95        23
    class_2       0.87      1.00      0.93        13

avg / total       0.95      0.94      0.95        36

But can I trust the model?

We can plot/display it directly

xgb.to_graphviz(xgb_clf_2, num_trees=0, size='100,150')

Feature Importance

features_importance = pd.DataFrame.from_dict(xgb_clf_2.get_fscore(), orient='index').reset_index().rename(columns={'index': 'features', 0: 'importance'})
features_importance.sort_values('importance', ascending=False, inplace=True)

features_importance.plot.barh(x='features', y='importance', figsize=(15,7))
plt.gca().invert_yaxis()

Partial Dependence Plot

• Display how response functions change based on the values of one or two independent variables of interest
• Averaging out the effects of all other independent variables

df_plot = pd.read_csv("./data/pdp/color_int.csv")
iplot(plotly_figure(df_plot, 'Color Intensity', 'Color Intensity',
                   'Probability of Class 2', 'color_intensity', "yhat"), 
     filename="figure")

df_plot = pd.read_csv("./data/pdp/flavanoids.csv")
iplot(plotly_figure(df_plot, 'Flavanoids', 'Flavanoids',
                   'Probability of Class 2', 'flavanoids', "yhat"), 
     filename="figure")

Local Interpretable Model-Agnostic Explanation (LIME)

• local linear approximation of the model's behaviour
• perturb the instance we want to explain and learn a sparse linear model around it

explainer = lime_tabular.LimeTabularExplainer(
    X_train, 
    feature_names=data.feature_names,
    class_names=data.target_names, 
    discretize_continuous=True
)

i = np.random.randint(0, X_test.shape[0])
print i
test_instance = X_test[i].copy()

0

exp = explainer.explain_instance(test_instance, xgb_clf.predict_proba, num_features=5)

exp.show_in_notebook(show_table=True, show_all=False)

i = np.random.randint(0, X_test.shape[0])
print i
test_instance = X_test[i].copy()

7

exp = explainer.explain_instance(test_instance, xgb_clf.predict_proba, num_features=5)

exp.show_in_notebook(show_table=True, show_all=False)

i = np.random.randint(0, X_test.shape[0])
print i
test_instance = X_test[i].copy()

18

exp = explainer.explain_instance(test_instance, xgb_clf.predict_proba, num_features=5)

exp.show_in_notebook(show_table=True, show_all=False)

test_instance[9] = 2.5

exp = explainer.explain_instance(test_instance, xgb_clf.predict_proba, num_features=5)

exp.show_in_notebook(show_table=True, show_all=False)

test_instance = X_test[i].copy()
test_instance[6] = 1.9

exp = explainer.explain_instance(test_instance, xgb_clf.predict_proba, num_features=5)

exp.show_in_notebook(show_table=True, show_all=False)

Interpreting Document Classification Model

Dataset used: 20 newsgroups

http://qwone.com/~jason/20Newsgroups/

The 20 Newsgroups data set is a collection of approximately 20,000 newsgroup documents, across 20 different newsgroups.

from sklearn.datasets import fetch_20newsgroups
newsgroups_train = fetch_20newsgroups(subset='train', remove=('headers', 'footers', 'quotes'))
newsgroups_test = fetch_20newsgroups(subset='test',remove=('headers', 'footers', 'quotes'))
# making class names shorter
class_names = [x.split('.')[-1] if 'misc' not in x else '.'.join(x.split('.')[-2:]) for x in newsgroups_train.target_names]
class_names[3] = 'pc.hardware'
class_names[4] = 'mac.hardware'

Downloading 20news dataset. This may take a few minutes.
Downloading dataset from https://ndownloader.figshare.com/files/5975967 (14 MB)

print(','.join(class_names))

atheism,graphics,ms-windows.misc,pc.hardware,mac.hardware,x,misc.forsale,autos,motorcycles,baseball,hockey,crypt,electronics,med,space,christian,guns,mideast,politics.misc,religion.misc

from sklearn.naive_bayes import MultinomialNB
from lime import lime_text
from sklearn.pipeline import make_pipeline
from lime.lime_text import LimeTextExplainer

train_vectors = vectorizer.fit_transform(newsgroups_train.data)
test_vectors = vectorizer.transform(newsgroups_test.data)
nb = MultinomialNB(alpha=.01)
nb.fit(train_vectors, newsgroups_train.target)
c = make_pipeline(vectorizer, nb)
explainer = LimeTextExplainer(class_names=class_names)

exp = explainer.explain_instance(newsgroups_test.data[idx], c.predict_proba, num_features=6, top_labels=2)
print(exp.available_labels())
exp.show_in_notebook(text=newsgroups_test.data[idx])

[8, 7]

exp = explainer.explain_instance(newsgroups_test.data[idx], c.predict_proba, num_features=6, top_labels=2)
print(exp.available_labels())
exp.show_in_notebook(text=newsgroups_test.data[idx])

[17, 15]

exp = explainer.explain_instance(newsgroups_test.data[idx], c.predict_proba, num_features=6, top_labels=2)
print(exp.available_labels())
exp.show_in_notebook(text=newsgroups_test.data[idx])

[12, 1]

Thank you

• https://github.com/marcotcr/lime
• https://www.oreilly.com/learning/introduction-to-local-interpretable-model-agnostic-explanations-lime
• https://www.oreilly.com/ideas/ideas-on-interpreting-machine-learning

Hiring at ds-jobs@go-jek.com!

Conducting Deep Learning Masterclass in NUS on 26th Oct together with 3 deep learning practitioners, for more information pls visit: NUS ISS Deep Learning Masterclass